AIDA: Artificial Intelligence Design Assistant

GDA proposes to build AIDA, a software system functioning as an Artificial Intelligence Design Assistant. To do this requires adapting and extending agent patterns from examples to novel settings. In particular, a design assistant for tasks such as naval vessel design must include the ability to process knowledge and interface with tools which are inherently multi-modal, including structured data, unstructured text, and images. Leveraging this expansion AIDA will include novel agent interfaces, where the assistant can propose, select, act, and observe effects of tools and knowledge queries. LLMs, because of their ability to use natural language input, have recently disrupted approaches to tasks that were previously thought to require human capability. This means technical skills such as programming will no longer be a barrier. However, experienced developers know that coding or programming is only part of the overall task of software engineering. As such, design and engineering tasks can be greatly assisted and accelerated by novel technologies, but it is critical to play to their strengths.  An important aspect of GenAI, one that leads to many issues, is the need to interpolate known solutions to obtain a return for a new one. This can lead to hallucinations, a ubiquitous issue for most all practical usage of GenAI. Funded by DARPA, GDA developed Topological Parallax as a safety specification for machine learning models, and will build this capability into AIDA.